CmmOpt.hs 27.4 KB
Newer Older
1 2 3 4 5 6 7
{-# OPTIONS -w #-}
-- The above warning supression flag is a temporary kludge.
-- While working on this module you are encouraged to remove it and fix
-- any warnings in the module. See
--     http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
-- for details

8 9 10 11 12 13 14 15 16
-----------------------------------------------------------------------------
--
-- Cmm optimisation
--
-- (c) The University of Glasgow 2006
--
-----------------------------------------------------------------------------

module CmmOpt (
17 18 19 20 21
        cmmEliminateDeadBlocks,
        cmmMiniInline,
        cmmMachOpFold,
        cmmMachOpFoldM,
        cmmLoopifyForC,
22 23 24 25
 ) where

#include "HsVersions.h"

26
import OldCmm
27
import CmmNode (wrapRecExp)
Simon Marlow's avatar
Simon Marlow committed
28 29
import CmmUtils
import CLabel
30
import StaticFlags
31 32

import UniqFM
Simon Marlow's avatar
Simon Marlow committed
33
import Unique
34
import FastTypes
35
import Outputable
36
import BlockId
37

Simon Marlow's avatar
Simon Marlow committed
38 39 40
import Data.Bits
import Data.Word
import Data.Int
41
import Data.Maybe
42
import Data.List
43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62

import Compiler.Hoopl hiding (Unique)

-- -----------------------------------------------------------------------------
-- Eliminates dead blocks

{-
We repeatedly expand the set of reachable blocks until we hit a
fixpoint, and then prune any blocks that were not in this set.  This is
actually a required optimization, as dead blocks can cause problems
for invariants in the linear register allocator (and possibly other
places.)
-}

-- Deep fold over statements could probably be abstracted out, but it
-- might not be worth the effort since OldCmm is moribund
cmmEliminateDeadBlocks :: [CmmBasicBlock] -> [CmmBasicBlock]
cmmEliminateDeadBlocks [] = []
cmmEliminateDeadBlocks blocks@(BasicBlock base_id _:_) =
    let -- Calculate what's reachable from what block
63 64 65
        reachableMap = foldl' f emptyUFM blocks -- lazy in values
            where f m (BasicBlock block_id stmts) = addToUFM m block_id (reachableFrom stmts)
        reachableFrom stmts = foldl stmt [] stmts
66 67 68 69 70 71 72 73
            where
                stmt m CmmNop = m
                stmt m (CmmComment _) = m
                stmt m (CmmAssign _ e) = expr m e
                stmt m (CmmStore e1 e2) = expr (expr m e1) e2
                stmt m (CmmCall c _ as _ _) = f (actuals m as) c
                    where f m (CmmCallee e _) = expr m e
                          f m (CmmPrim _) = m
74 75 76
                stmt m (CmmBranch b) = b:m
                stmt m (CmmCondBranch e b) = b:(expr m e)
                stmt m (CmmSwitch e bs) = catMaybes bs ++ expr m e
77 78
                stmt m (CmmJump e as) = expr (actuals m as) e
                stmt m (CmmReturn as) = actuals m as
79 80 81
                actuals m as = foldl' (\m h -> expr m (hintlessCmm h)) m as
                -- We have to do a deep fold into CmmExpr because
                -- there may be a BlockId in the CmmBlock literal.
82 83 84
                expr m (CmmLit l) = lit m l
                expr m (CmmLoad e _) = expr m e
                expr m (CmmReg _) = m
85
                expr m (CmmMachOp _ es) = foldl' expr m es
86 87
                expr m (CmmStackSlot _ _) = m
                expr m (CmmRegOff _ _) = m
88
                lit m (CmmBlock b) = b:m
89
                lit m _ = m
90 91 92 93 94 95 96 97
        -- go todo done
        reachable = go [base_id] (setEmpty :: BlockSet)
          where go []     m = m
                go (x:xs) m
                    | setMember x m = go xs m
                    | otherwise     = go (add ++ xs) (setInsert x m)
                        where add = fromMaybe (panic "cmmEliminateDeadBlocks: unknown block")
                                              (lookupUFM reachableMap x)
98
    in filter (\(BasicBlock block_id _) -> setMember block_id reachable) blocks
99 100 101 102

-- -----------------------------------------------------------------------------
-- The mini-inliner

Simon Marlow's avatar
Simon Marlow committed
103
{-
104 105 106 107
This pass inlines assignments to temporaries.  Temporaries that are
only used once are unconditionally inlined.  Temporaries that are used
two or more times are only inlined if they are assigned a literal.  It
works as follows:
Simon Marlow's avatar
Simon Marlow committed
108 109

  - count uses of each temporary
110
  - for each temporary:
Simon Marlow's avatar
Simon Marlow committed
111 112 113 114 115 116
	- attempt to push it forward to the statement that uses it
        - only push forward past assignments to other temporaries
	  (assumes that temporaries are single-assignment)
	- if we reach the statement that uses it, inline the rhs
	  and delete the original assignment.

117 118 119 120
[N.B. In the Quick C-- compiler, this optimization is achieved by a
 combination of two dataflow passes: forward substitution (peephole
 optimization) and dead-assignment elimination.  ---NR]

Simon Marlow's avatar
Simon Marlow committed
121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
Possible generalisations: here is an example from factorial

Fac_zdwfac_entry:
    cmG:
        _smi = R2;
        if (_smi != 0) goto cmK;
        R1 = R3;
        jump I64[Sp];
    cmK:
        _smn = _smi * R3;
        R2 = _smi + (-1);
        R3 = _smn;
        jump Fac_zdwfac_info;

We want to inline _smi and _smn.  To inline _smn:

   - we must be able to push forward past assignments to global regs.
     We can do this if the rhs of the assignment we are pushing
     forward doesn't refer to the global reg being assigned to; easy
     to test.

To inline _smi:

   - It is a trivial replacement, reg for reg, but it occurs more than
     once.
   - We can inline trivial assignments even if the temporary occurs
     more than once, as long as we don't eliminate the original assignment
     (this doesn't help much on its own).
   - We need to be able to propagate the assignment forward through jumps;
     if we did this, we would find that it can be inlined safely in all
     its occurrences.
-}

154 155 156 157
countUses :: UserOfLocalRegs a => a -> UniqFM Int
countUses a = foldRegsUsed (\m r -> addToUFM m r (count m r + 1)) emptyUFM a
  where count m r = lookupWithDefaultUFM m (0::Int) r

158 159
cmmMiniInline :: [CmmBasicBlock] -> [CmmBasicBlock]
cmmMiniInline blocks = map do_inline blocks 
160 161
  where do_inline (BasicBlock id stmts)
          = BasicBlock id (cmmMiniInlineStmts (countUses blocks) stmts)
162 163 164

cmmMiniInlineStmts :: UniqFM Int -> [CmmStmt] -> [CmmStmt]
cmmMiniInlineStmts uses [] = []
165
cmmMiniInlineStmts uses (stmt@(CmmAssign (CmmLocal (LocalReg u _)) expr) : stmts)
166 167 168 169
        -- not used: just discard this assignment
  | Nothing <- lookupUFM uses u
  = cmmMiniInlineStmts uses stmts

170 171
        -- used (literal): try to inline at all the use sites
  | Just n <- lookupUFM uses u, isLit expr
172 173 174 175 176 177 178 179 180 181
  =
#ifdef NCG_DEBUG
     trace ("nativeGen: inlining " ++ showSDoc (pprStmt stmt)) $
#endif
     case lookForInlineLit u expr stmts of
         (m, stmts')
             | n == m -> cmmMiniInlineStmts (delFromUFM uses u) stmts'
             | otherwise ->
                 stmt : cmmMiniInlineStmts (adjustUFM (\x -> x - m) uses u) stmts'

182 183
        -- used (foldable to literal): try to inline at all the use sites
  | Just n <- lookupUFM uses u,
184
    e@(CmmLit _) <- wrapRecExp foldExp expr
185 186 187 188 189 190 191 192 193
  =
#ifdef NCG_DEBUG
     trace ("nativeGen: inlining " ++ showSDoc (pprStmt stmt)) $
#endif
     case lookForInlineLit u e stmts of
         (m, stmts')
             | n == m -> cmmMiniInlineStmts (delFromUFM uses u) stmts'
             | otherwise ->
                 stmt : cmmMiniInlineStmts (adjustUFM (\x -> x - m) uses u) stmts'
194

195
        -- used once (non-literal): try to inline at the use site
196 197 198 199 200 201 202
  | Just 1 <- lookupUFM uses u,
    Just stmts' <- lookForInline u expr stmts
  = 
#ifdef NCG_DEBUG
     trace ("nativeGen: inlining " ++ showSDoc (pprStmt stmt)) $
#endif
     cmmMiniInlineStmts uses stmts'
203 204 205
 where
  foldExp (CmmMachOp op args) = cmmMachOpFold op args
  foldExp e = e
206 207 208 209

cmmMiniInlineStmts uses (stmt:stmts)
  = stmt : cmmMiniInlineStmts uses stmts

210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234
-- | Takes a register, a 'CmmLit' expression assigned to that
-- register, and a list of statements.  Inlines the expression at all
-- use sites of the register.  Returns the number of substituations
-- made and the, possibly modified, list of statements.
lookForInlineLit :: Unique -> CmmExpr -> [CmmStmt] -> (Int, [CmmStmt])
lookForInlineLit _ _ [] = (0, [])
lookForInlineLit u expr stmts@(stmt : rest)
  | Just n <- lookupUFM (countUses stmt) u
  = case lookForInlineLit u expr rest of
      (m, stmts) -> let z = n + m
                    in z `seq` (z, inlineStmt u expr stmt : stmts)

  | ok_to_skip
  = case lookForInlineLit u expr rest of
      (n, stmts) -> (n, stmt : stmts)

  | otherwise
  = (0, stmts)
  where
    -- We skip over assignments to registers, unless the register
    -- being assigned to is the one we're inlining.
    ok_to_skip = case stmt of
        CmmAssign (CmmLocal r@(LocalReg u' _)) _ | u' == u -> False
        _other -> True

235 236 237 238
lookForInline u expr stmts = lookForInline' u expr regset stmts
    where regset = foldRegsUsed extendRegSet emptyRegSet expr

lookForInline' u expr regset (stmt : rest)
239 240 241 242
  | Just 1 <- lookupUFM (countUses stmt) u, ok_to_inline
  = Just (inlineStmt u expr stmt : rest)

  | ok_to_skip
243
  = case lookForInline' u expr regset rest of
244 245 246 247 248
           Nothing    -> Nothing
           Just stmts -> Just (stmt:stmts)

  | otherwise 
  = Nothing
249

250 251 252 253 254 255 256 257 258
  where
	-- we don't inline into CmmCall if the expression refers to global
	-- registers.  This is a HACK to avoid global registers clashing with
	-- C argument-passing registers, really the back-end ought to be able
	-- to handle it properly, but currently neither PprC nor the NCG can
	-- do it.  See also CgForeignCall:load_args_into_temps.
    ok_to_inline = case stmt of
		     CmmCall{} -> hasNoGlobalRegs expr
		     _ -> True
259

260 261 262 263 264 265 266 267
   -- Expressions aren't side-effecting.  Temporaries may or may not
   -- be single-assignment depending on the source (the old code
   -- generator creates single-assignment code, but hand-written Cmm
   -- and Cmm from the new code generator is not single-assignment.)
   -- So we do an extra check to make sure that the register being
   -- changed is not one we were relying on.  I don't know how much of a
   -- performance hit this is (we have to create a regset for every
   -- instruction.) -- EZY
268 269
    ok_to_skip = case stmt of
                 CmmNop -> True
270
                 CmmComment{} -> True
271
                 CmmAssign (CmmLocal r@(LocalReg u' _)) rhs | u' /= u && not (r `elemRegSet` regset) -> True
272 273 274 275
                 CmmAssign g@(CmmGlobal _) rhs -> not (g `regUsedIn` expr)
                 _other -> False


276 277 278
inlineStmt :: Unique -> CmmExpr -> CmmStmt -> CmmStmt
inlineStmt u a (CmmAssign r e) = CmmAssign r (inlineExpr u a e)
inlineStmt u a (CmmStore e1 e2) = CmmStore (inlineExpr u a e1) (inlineExpr u a e2)
279 280
inlineStmt u a (CmmCall target regs es srt ret)
   = CmmCall (infn target) regs es' srt ret
281
   where infn (CmmCallee fn cconv) = CmmCallee (inlineExpr u a fn) cconv
282
	 infn (CmmPrim p) = CmmPrim p
283
	 es' = [ (CmmHinted (inlineExpr u a e) hint) | (CmmHinted e hint) <- es ]
284 285 286 287 288 289
inlineStmt u a (CmmCondBranch e d) = CmmCondBranch (inlineExpr u a e) d
inlineStmt u a (CmmSwitch e d) = CmmSwitch (inlineExpr u a e) d
inlineStmt u a (CmmJump e d) = CmmJump (inlineExpr u a e) d
inlineStmt u a other_stmt = other_stmt

inlineExpr :: Unique -> CmmExpr -> CmmExpr -> CmmExpr
290
inlineExpr u a e@(CmmReg (CmmLocal (LocalReg u' _)))
291 292
  | u == u' = a
  | otherwise = e
293 294
inlineExpr u a e@(CmmRegOff (CmmLocal (LocalReg u' rep)) off)
  | u == u' = CmmMachOp (MO_Add width) [a, CmmLit (CmmInt (fromIntegral off) width)]
295
  | otherwise = e
296 297
  where
    width = typeWidth rep
298 299 300 301 302 303 304 305 306 307 308
inlineExpr u a (CmmLoad e rep) = CmmLoad (inlineExpr u a e) rep
inlineExpr u a (CmmMachOp op es) = CmmMachOp op (map (inlineExpr u a) es)
inlineExpr u a other_expr = other_expr

-- -----------------------------------------------------------------------------
-- MachOp constant folder

-- Now, try to constant-fold the MachOps.  The arguments have already
-- been optimized and folded.

cmmMachOpFold
309 310
    :: MachOp       -- The operation from an CmmMachOp
    -> [CmmExpr]    -- The optimized arguments
311 312
    -> CmmExpr

313 314 315 316 317 318 319 320 321 322 323
cmmMachOpFold op args = fromMaybe (CmmMachOp op args) (cmmMachOpFoldM op args)

-- Returns Nothing if no changes, useful for Hoopl, also reduces
-- allocation!
cmmMachOpFoldM
    :: MachOp
    -> [CmmExpr]
    -> Maybe CmmExpr

cmmMachOpFoldM op arg@[CmmLit (CmmInt x rep)]
  = Just $ case op of
324 325 326
      MO_S_Neg r -> CmmLit (CmmInt (-x) rep)
      MO_Not r   -> CmmLit (CmmInt (complement x) rep)

327 328 329 330
        -- these are interesting: we must first narrow to the 
        -- "from" type, in order to truncate to the correct size.
        -- The final narrow/widen to the destination type
        -- is implicit in the CmmLit.
331 332 333
      MO_SF_Conv from to -> CmmLit (CmmFloat (fromInteger x) to)
      MO_SS_Conv from to -> CmmLit (CmmInt (narrowS from x) to)
      MO_UU_Conv from to -> CmmLit (CmmInt (narrowU from x) to)
334

335
      _ -> panic "cmmMachOpFoldM: unknown unary op"
336 337 338


-- Eliminate conversion NOPs
339 340
cmmMachOpFoldM (MO_SS_Conv rep1 rep2) [x] | rep1 == rep2 = Just x
cmmMachOpFoldM (MO_UU_Conv rep1 rep2) [x] | rep1 == rep2 = Just x
341 342

-- Eliminate nested conversions where possible
343
cmmMachOpFoldM conv_outer args@[CmmMachOp conv_inner [x]]
344 345 346
  | Just (rep1,rep2,signed1) <- isIntConversion conv_inner,
    Just (_,   rep3,signed2) <- isIntConversion conv_outer
  = case () of
347 348 349 350 351 352 353 354 355 356 357 358 359 360 361
        -- widen then narrow to the same size is a nop
      _ | rep1 < rep2 && rep1 == rep3 -> Just x
        -- Widen then narrow to different size: collapse to single conversion
        -- but remember to use the signedness from the widening, just in case
        -- the final conversion is a widen.
        | rep1 < rep2 && rep2 > rep3 ->
            Just $ cmmMachOpFold (intconv signed1 rep1 rep3) [x]
        -- Nested widenings: collapse if the signedness is the same
        | rep1 < rep2 && rep2 < rep3 && signed1 == signed2 ->
            Just $ cmmMachOpFold (intconv signed1 rep1 rep3) [x]
        -- Nested narrowings: collapse
        | rep1 > rep2 && rep2 > rep3 ->
            Just $ cmmMachOpFold (MO_UU_Conv rep1 rep3) [x]
        | otherwise ->
            Nothing
362
  where
363 364 365 366 367
        isIntConversion (MO_UU_Conv rep1 rep2) 
          = Just (rep1,rep2,False)
        isIntConversion (MO_SS_Conv rep1 rep2)
          = Just (rep1,rep2,True)
        isIntConversion _ = Nothing
368

369 370
        intconv True  = MO_SS_Conv
        intconv False = MO_UU_Conv
371 372 373 374 375

-- ToDo: a narrow of a load can be collapsed into a narrow load, right?
-- but what if the architecture only supports word-sized loads, should
-- we do the transformation anyway?

376
cmmMachOpFoldM mop args@[CmmLit (CmmInt x xrep), CmmLit (CmmInt y _)]
377
  = case mop of
378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409
        -- for comparisons: don't forget to narrow the arguments before
        -- comparing, since they might be out of range.
        MO_Eq r   -> Just $ CmmLit (CmmInt (if x_u == y_u then 1 else 0) wordWidth)
        MO_Ne r   -> Just $ CmmLit (CmmInt (if x_u /= y_u then 1 else 0) wordWidth)

        MO_U_Gt r -> Just $ CmmLit (CmmInt (if x_u >  y_u then 1 else 0) wordWidth)
        MO_U_Ge r -> Just $ CmmLit (CmmInt (if x_u >= y_u then 1 else 0) wordWidth)
        MO_U_Lt r -> Just $ CmmLit (CmmInt (if x_u <  y_u then 1 else 0) wordWidth)
        MO_U_Le r -> Just $ CmmLit (CmmInt (if x_u <= y_u then 1 else 0) wordWidth)

        MO_S_Gt r -> Just $ CmmLit (CmmInt (if x_s >  y_s then 1 else 0) wordWidth)
        MO_S_Ge r -> Just $ CmmLit (CmmInt (if x_s >= y_s then 1 else 0) wordWidth)
        MO_S_Lt r -> Just $ CmmLit (CmmInt (if x_s <  y_s then 1 else 0) wordWidth)
        MO_S_Le r -> Just $ CmmLit (CmmInt (if x_s <= y_s then 1 else 0) wordWidth)

        MO_Add r -> Just $ CmmLit (CmmInt (x + y) r)
        MO_Sub r -> Just $ CmmLit (CmmInt (x - y) r)
        MO_Mul r -> Just $ CmmLit (CmmInt (x * y) r)
        MO_U_Quot r | y /= 0 -> Just $ CmmLit (CmmInt (x_u `quot` y_u) r)
        MO_U_Rem  r | y /= 0 -> Just $ CmmLit (CmmInt (x_u `rem`  y_u) r)
        MO_S_Quot r | y /= 0 -> Just $ CmmLit (CmmInt (x `quot` y) r)
        MO_S_Rem  r | y /= 0 -> Just $ CmmLit (CmmInt (x `rem` y) r)

        MO_And   r -> Just $ CmmLit (CmmInt (x .&. y) r)
        MO_Or    r -> Just $ CmmLit (CmmInt (x .|. y) r)
        MO_Xor   r -> Just $ CmmLit (CmmInt (x `xor` y) r)

        MO_Shl   r -> Just $ CmmLit (CmmInt (x `shiftL` fromIntegral y) r)
        MO_U_Shr r -> Just $ CmmLit (CmmInt (x_u `shiftR` fromIntegral y) r)
        MO_S_Shr r -> Just $ CmmLit (CmmInt (x `shiftR` fromIntegral y) r)

        other      -> Nothing
410 411

   where
412 413 414 415 416
        x_u = narrowU xrep x
        y_u = narrowU xrep y
        x_s = narrowS xrep x
        y_s = narrowS xrep y

417 418 419 420 421 422

-- When possible, shift the constants to the right-hand side, so that we
-- can match for strength reductions.  Note that the code generator will
-- also assume that constants have been shifted to the right when
-- possible.

423 424 425
cmmMachOpFoldM op [x@(CmmLit _), y]
   | not (isLit y) && isCommutableMachOp op
   = Just (cmmMachOpFold op [y, x])
426 427 428 429 430 431 432 433 434 435 436 437 438 439

-- Turn (a+b)+c into a+(b+c) where possible.  Because literals are
-- moved to the right, it is more likely that we will find
-- opportunities for constant folding when the expression is
-- right-associated.
--
-- ToDo: this appears to introduce a quadratic behaviour due to the
-- nested cmmMachOpFold.  Can we fix this?
--
-- Why do we check isLit arg1?  If arg1 is a lit, it means that arg2
-- is also a lit (otherwise arg1 would be on the right).  If we
-- put arg1 on the left of the rearranged expression, we'll get into a
-- loop:  (x1+x2)+x3 => x1+(x2+x3)  => (x2+x3)+x1 => x2+(x3+x1) ...
--
440 441 442
-- Also don't do it if arg1 is PicBaseReg, so that we don't separate the
-- PicBaseReg from the corresponding label (or label difference).
--
443
cmmMachOpFoldM mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
444
   | mop2 `associates_with` mop1
445
     && not (isLit arg1) && not (isPicReg arg1)
446
   = Just (cmmMachOpFold mop2 [arg1, cmmMachOpFold mop1 [arg2,arg3]])
447 448 449 450 451 452
   where
     MO_Add{} `associates_with` MO_Sub{} = True
     mop1 `associates_with` mop2 =
        mop1 == mop2 && isAssociativeMachOp mop1

-- special case: (a - b) + c  ==>  a + (c - b)
453
cmmMachOpFoldM mop1@(MO_Add{}) [CmmMachOp mop2@(MO_Sub{}) [arg1,arg2], arg3]
454
   | not (isLit arg1) && not (isPicReg arg1)
455
   = Just (cmmMachOpFold mop1 [arg1, cmmMachOpFold mop2 [arg3,arg2]])
456 457

-- Make a RegOff if we can
458 459 460 461 462 463 464 465
cmmMachOpFoldM (MO_Add _) [CmmReg reg, CmmLit (CmmInt n rep)]
  = Just $ CmmRegOff reg (fromIntegral (narrowS rep n))
cmmMachOpFoldM (MO_Add _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
  = Just $ CmmRegOff reg (off + fromIntegral (narrowS rep n))
cmmMachOpFoldM (MO_Sub _) [CmmReg reg, CmmLit (CmmInt n rep)]
  = Just $ CmmRegOff reg (- fromIntegral (narrowS rep n))
cmmMachOpFoldM (MO_Sub _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
  = Just $ CmmRegOff reg (off - fromIntegral (narrowS rep n))
466 467 468

-- Fold label(+/-)offset into a CmmLit where possible

469 470 471 472 473 474
cmmMachOpFoldM (MO_Add _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
  = Just $ CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
cmmMachOpFoldM (MO_Add _) [CmmLit (CmmInt i rep), CmmLit (CmmLabel lbl)]
  = Just $ CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
cmmMachOpFoldM (MO_Sub _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
  = Just $ CmmLit (CmmLabelOff lbl (fromIntegral (negate (narrowU rep i))))
475

476

477 478 479 480 481 482
-- Comparison of literal with widened operand: perform the comparison
-- at the smaller width, as long as the literal is within range.

-- We can't do the reverse trick, when the operand is narrowed:
-- narrowing throws away bits from the operand, there's no way to do
-- the same comparison at the larger size.
483 484 485 486

#if i386_TARGET_ARCH || x86_64_TARGET_ARCH
-- powerPC NCG has a TODO for I8/I16 comparisons, so don't try

487
cmmMachOpFoldM cmp [CmmMachOp conv [x], CmmLit (CmmInt i _)]
488 489 490 491 492 493 494
  |     -- if the operand is widened:
    Just (rep, signed, narrow_fn) <- maybe_conversion conv,
        -- and this is a comparison operation:
    Just narrow_cmp <- maybe_comparison cmp rep signed,
        -- and the literal fits in the smaller size:
    i == narrow_fn rep i
        -- then we can do the comparison at the smaller size
495
  = Just (cmmMachOpFold narrow_cmp [x, CmmLit (CmmInt i rep)])
496
 where
497
    maybe_conversion (MO_UU_Conv from to)
498 499
        | to > from
        = Just (from, False, narrowU)
500 501
    maybe_conversion (MO_SS_Conv from to)
        | to > from
502
        = Just (from, True, narrowS)
503

Simon Marlow's avatar
Simon Marlow committed
504 505
        -- don't attempt to apply this optimisation when the source
        -- is a float; see #1916
506
    maybe_conversion _ = Nothing
507

508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
        -- careful (#2080): if the original comparison was signed, but
        -- we were doing an unsigned widen, then we must do an
        -- unsigned comparison at the smaller size.
    maybe_comparison (MO_U_Gt _) rep _     = Just (MO_U_Gt rep)
    maybe_comparison (MO_U_Ge _) rep _     = Just (MO_U_Ge rep)
    maybe_comparison (MO_U_Lt _) rep _     = Just (MO_U_Lt rep)
    maybe_comparison (MO_U_Le _) rep _     = Just (MO_U_Le rep)
    maybe_comparison (MO_Eq   _) rep _     = Just (MO_Eq   rep)
    maybe_comparison (MO_S_Gt _) rep True  = Just (MO_S_Gt rep)
    maybe_comparison (MO_S_Ge _) rep True  = Just (MO_S_Ge rep)
    maybe_comparison (MO_S_Lt _) rep True  = Just (MO_S_Lt rep)
    maybe_comparison (MO_S_Le _) rep True  = Just (MO_S_Le rep)
    maybe_comparison (MO_S_Gt _) rep False = Just (MO_U_Gt rep)
    maybe_comparison (MO_S_Ge _) rep False = Just (MO_U_Ge rep)
    maybe_comparison (MO_S_Lt _) rep False = Just (MO_U_Lt rep)
    maybe_comparison (MO_S_Le _) rep False = Just (MO_U_Le rep)
    maybe_comparison _ _ _ = Nothing
525 526 527

#endif

528 529
-- We can often do something with constants of 0 and 1 ...

530
cmmMachOpFoldM mop args@[x, y@(CmmLit (CmmInt 0 _))]
531
  = case mop of
532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553
        MO_Add   r -> Just x
        MO_Sub   r -> Just x
        MO_Mul   r -> Just y
        MO_And   r -> Just y
        MO_Or    r -> Just x
        MO_Xor   r -> Just x
        MO_Shl   r -> Just x
        MO_S_Shr r -> Just x
        MO_U_Shr r -> Just x
        MO_Ne    r | isComparisonExpr x -> Just x
        MO_Eq    r | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_U_Gt  r | isComparisonExpr x -> Just x
        MO_S_Gt  r | isComparisonExpr x -> Just x
        MO_U_Lt  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_S_Lt  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_U_Ge  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_S_Ge  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_U_Le  r | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_S_Le  r | Just x' <- maybeInvertCmmExpr x -> Just x'
        other    -> Nothing

cmmMachOpFoldM mop args@[x, y@(CmmLit (CmmInt 1 rep))]
554
  = case mop of
555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570
        MO_Mul    r -> Just x
        MO_S_Quot r -> Just x
        MO_U_Quot r -> Just x
        MO_S_Rem  r -> Just $ CmmLit (CmmInt 0 rep)
        MO_U_Rem  r -> Just $ CmmLit (CmmInt 0 rep)
        MO_Ne    r | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_Eq    r | isComparisonExpr x -> Just x
        MO_U_Lt  r | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_S_Lt  r | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_U_Gt  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_S_Gt  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_U_Le  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_S_Le  r | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_U_Ge  r | isComparisonExpr x -> Just x
        MO_S_Ge  r | isComparisonExpr x -> Just x
        other       -> Nothing
571 572 573

-- Now look for multiplication/division by powers of 2 (integers).

574
cmmMachOpFoldM mop args@[x, y@(CmmLit (CmmInt n _))]
575
  = case mop of
576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611
        MO_Mul rep
           | Just p <- exactLog2 n ->
                 Just (cmmMachOpFold (MO_Shl rep) [x, CmmLit (CmmInt p rep)])
        MO_U_Quot rep
           | Just p <- exactLog2 n ->
                 Just (cmmMachOpFold (MO_U_Shr rep) [x, CmmLit (CmmInt p rep)])
        MO_S_Quot rep
           | Just p <- exactLog2 n, 
             CmmReg _ <- x ->   -- We duplicate x below, hence require
                                -- it is a reg.  FIXME: remove this restriction.
                -- shift right is not the same as quot, because it rounds
                -- to minus infinity, whereasq quot rounds toward zero.
                -- To fix this up, we add one less than the divisor to the
                -- dividend if it is a negative number.
                --
                -- to avoid a test/jump, we use the following sequence:
                -- 	x1 = x >> word_size-1  (all 1s if -ve, all 0s if +ve)
                --      x2 = y & (divisor-1)
                --      result = (x+x2) >>= log2(divisor)
                -- this could be done a bit more simply using conditional moves,
                -- but we're processor independent here.
                --
                -- we optimise the divide by 2 case slightly, generating
                --      x1 = x >> word_size-1  (unsigned)
                --      return = (x + x1) >>= log2(divisor)
                let
                    bits = fromIntegral (widthInBits rep) - 1
                    shr = if p == 1 then MO_U_Shr rep else MO_S_Shr rep
                    x1 = CmmMachOp shr [x, CmmLit (CmmInt bits rep)]
                    x2 = if p == 1 then x1 else
                         CmmMachOp (MO_And rep) [x1, CmmLit (CmmInt (n-1) rep)]
                    x3 = CmmMachOp (MO_Add rep) [x, x2]
                in
                Just (cmmMachOpFold (MO_S_Shr rep) [x3, CmmLit (CmmInt p rep)])
        other
           -> Nothing
612 613 614

-- Anything else is just too hard.

615
cmmMachOpFoldM _ _ = Nothing
616 617 618 619 620 621 622 623 624 625 626

-- -----------------------------------------------------------------------------
-- exactLog2

-- This algorithm for determining the $\log_2$ of exact powers of 2 comes
-- from GCC.  It requires bit manipulation primitives, and we use GHC
-- extensions.  Tough.
-- 
-- Used to be in MachInstrs --SDM.
-- ToDo: remove use of unboxery --SDM.

627 628 629 630 631
-- Unboxery removed in favor of FastInt; but is the function supposed to fail
-- on inputs >= 2147483648, or was that just an implementation artifact?
-- And is this speed-critical, or can we just use Integer operations
-- (including Data.Bits)?
--  --Isaac Dupree
632 633

exactLog2 :: Integer -> Maybe Integer
634 635
exactLog2 x_
  = if (x_ <= 0 || x_ >= 2147483648) then
636 637
       Nothing
    else
638 639
       case iUnbox (fromInteger x_) of { x ->
       if (x `bitAndFastInt` negateFastInt x) /=# x then
640 641
	  Nothing
       else
642
	  Just (toInteger (iBox (pow2 x)))
643 644
       }
  where
645 646
    pow2 x | x ==# _ILIT(1) = _ILIT(0)
           | otherwise = _ILIT(1) +# pow2 (x `shiftR_FastInt` _ILIT(1))
647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674


-- -----------------------------------------------------------------------------
-- Loopify for C

{-
 This is a simple pass that replaces tail-recursive functions like this:

   fac() {
     ...
     jump fac();
   }

 with this:

  fac() {
   L:
     ...
     goto L;
  }

  the latter generates better C code, because the C compiler treats it
  like a loop, and brings full loop optimisation to bear.

  In my measurements this makes little or no difference to anything
  except factorial, but what the hell.
-}

675
cmmLoopifyForC :: RawCmmTop -> RawCmmTop
676
cmmLoopifyForC p@(CmmProc info entry_lbl
677
                 (ListGraph blocks@(BasicBlock top_id _ : _)))
678 679 680
  | null info = p  -- only if there's an info table, ignore case alts
  | otherwise =  
--  pprTrace "jump_lbl" (ppr jump_lbl <+> ppr entry_lbl) $
681
  CmmProc info entry_lbl (ListGraph blocks')
682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703
  where blocks' = [ BasicBlock id (map do_stmt stmts)
		  | BasicBlock id stmts <- blocks ]

        do_stmt (CmmJump (CmmLit (CmmLabel lbl)) _) | lbl == jump_lbl
		= CmmBranch top_id
	do_stmt stmt = stmt

	jump_lbl | tablesNextToCode = entryLblToInfoLbl entry_lbl
		 | otherwise        = entry_lbl

cmmLoopifyForC top = top

-- -----------------------------------------------------------------------------
-- Utils

isLit (CmmLit _) = True
isLit _          = False

isComparisonExpr :: CmmExpr -> Bool
isComparisonExpr (CmmMachOp op _) = isComparisonMachOp op
isComparisonExpr _other 	    = False

704
isPicReg (CmmReg (CmmGlobal PicBaseReg)) = True
705
isPicReg _ = False
706